Measured comparison · August 2026
Claude Code security review vs Ozone
Ozone caught 63 of 82; Claude Code caught 35. Both are careful when they speak — Claude Code raised 10 false alarms, Ozone 23 — but Claude Code missed 47 of the vulnerabilities it was shown.
The important caveat: Claude Code was run as a CLI against the filesystem, not as a pull request review, and in batches of files rather than as a diff. That is not how a reviewer sees a change, and it is likely part of why the number is what it is.
The honest framing is that a general-purpose coding agent given a security prompt is not the same thing as a pipeline built for review, and this measures the gap rather than the model.
| Tool | Vulnerabilities caught | False alarms | |
|---|---|---|---|
| Claude Code | 35 of 82 | 10 | |
| Ozone | 63 of 82 | 23 |
OpenSSF CVE Benchmark, 82 real JavaScript and TypeScript vulnerabilities, each also shown patched to count false alarms. Competitors' findings are their own published output (DeepSource, April 2026); the grading is ours, applied identically to all ten and blind to which tool produced a finding.
Different tools wearing similar labels
Claude Code is an agentic coding assistant with a `/security-review` command. It is genuinely capable, and 35 of 82 with only 10 false alarms is a respectable result for something that is not primarily a review product.
Ozone is a review pipeline: it clones the repository, builds a model of how the code fits together, reviews the diff in the context of everything it calls and everything that calls it, then reports on the pull request before merge. The infrastructure around the model is most of the product.
Why context changes the number
The vulnerabilities this benchmark uses are largely not visible in a single file. Prototype pollution, path traversal and command injection are reachable only if attacker-controlled input can get to the sink, and answering that means following callers across the codebase.
A batched file scan cannot do that reliably. Neither can a diff read in isolation. This is the specific thing a review pipeline is built to do, and it accounts for much of the 63-versus-35 difference.
Where the comparison is unfair, stated plainly
Claude Code's figures come from DeepSource's April 2026 run, in CLI mode, batching ten files per invocation because larger batches degraded. Ours come from Ozone's ordinary pull request path in August 2026.
There is a further reason to treat its score as a floor. Across the run Claude Code reported nothing at all on 93 of 165 files — by far the most of any tool tested, and a pattern more consistent with a degraded harness than with a reviewer that looked and found nothing. For contrast, CodeRabbit was silent on 37 and produced 376 findings in total. We would not lean on this comparison the way we would lean on that one.
Those are not equivalent conditions and we would not want the number quoted as though they were. Anthropic has since shipped a dedicated security product, which this benchmark does not cover at all.
When to choose Claude Code
- You want one tool that writes code, refactors, runs tests and can also review — rather than a dedicated reviewer sitting in CI.
- You value precision over coverage and are comfortable that roughly half the vulnerabilities will not be flagged. 10 false alarms across 82 patched files is a clean result.
- You are already paying for it, and marginal security review is a bonus rather than the purchase.
- Its score here was produced outside its native mode and outside a pull request. Treat 35 of 82 as a floor.
Common questions
- Can Claude Code find security vulnerabilities?
- Yes — it caught 35 of 82 real vulnerabilities in the OpenSSF CVE Benchmark, with 10 false alarms across the patched versions. It was run as a CLI over batched files rather than as a pull request review, so the figure is likely understated.
- Do I need a dedicated code review tool if I use Claude Code?
- It depends on coverage. Claude Code missed 47 of 82 vulnerabilities in this run against Ozone's 19. If a missed vulnerability reaching production is expensive for you, a dedicated review pipeline measured meaningfully better.
Check the numbers yourself.
The harness, the fixtures, every raw finding and every grader verdict are published. Re-run it and tell us if you get something different.